Time frame based ITS tracking

[mpuccio]

Hi @shahor02 @iouribelikov,
I recovered quite some efficiency but I am still investigating a random crash. I open this in draft while I fix that.
Cheers,
Max

[mpuccio]

Hi @shahor02 @iouribelikov , this is finally ready to be reviewed/merged. I still have to add the bit for flagging the tracks with clusters on more than one ROF, but this should be faster than this large modification.

[mpuccio]

@mconcas the last commit should fix the compilation of the ALICE3 macro. However, the functionality is not there: the tf are not properly filled with clusters. One should add an event by event fill method for the time frame to make it work

[mpuccio]

Hi @iouribelikov, now one can disable the multi-rof reconstruction by setting deltaROF=0 in the configuration of the tracker. If fine with you I would merge this, such that with Matteo we can synchronise the next developments

[iouribelikov]

Hi @iouribelikov, now one can disable the multi-rof reconstruction by setting deltaROF=0 in the configuration of the tracker. If fine with you I would merge this, such that with Matteo we can synchronise the next developments

Ciao, @mpuccio. In general, I do not have anything against merging these two PRs. It's just I do not see much improvement in the efficiency and fake-track rate.

Can we get a "snapshot" of the current performance (efficiency, fakes, CPU, memory), with the current best values for all tracking parameters ? I think, it is important to be sure that we at least do not degrade our performance, after the merging is done.

[mpuccio]

OK, for pp using your 1000 pp events with this PR I get:

	Command being timed: "o2-its-reco-workflow --trackerCA --tracking-mode async --configKeyValues ITSVertexerParam.phiCut=0.5;ITSVertexerParam.clusterContributorsCut=3;ITSVertexerParam.tanLambdaCut=0.2 -b"
	User time (seconds): 9.91
	System time (seconds): 2.83
	Percent of CPU this job got: 241%
	Elapsed (wall clock) time (h:mm:ss or m:ss): 0:05.28
	Average shared text size (kbytes): 0
	Average unshared data size (kbytes): 0
	Average stack size (kbytes): 0
	Average total size (kbytes): 0
	Maximum resident set size (kbytes): 2734340
	Average resident set size (kbytes): 0
	Major (requiring I/O) page faults: 0
	Minor (reclaiming a frame) page faults: 1110460
	Voluntary context switches: 33931
	Involuntary context switches: 1304
	Swaps: 0
	File system inputs: 0
	File system outputs: 42768
	Socket messages sent: 0
	Socket messages received: 0
	Signals delivered: 0
	Page size (bytes): 4096
	Exit status: 0

Corresponding to:
Good found tracks/event: 4, efficiency: 0.939128, fake-track rate: 0.0675035 clone rate 0.0168167

With the current dev I get:

	Command being timed: "o2-its-reco-workflow --trackerCA --tracking-mode async --configKeyValues ITSVertexerParam.phiCut=0.5;ITSVertexerParam.clusterContributorsCut=3;ITSVertexerParam.tanLambdaCut=0.2 -b"
	User time (seconds): 8.50
	System time (seconds): 2.31
	Percent of CPU this job got: 251%
	Elapsed (wall clock) time (h:mm:ss or m:ss): 0:04.29
	Average shared text size (kbytes): 0
	Average unshared data size (kbytes): 0
	Average stack size (kbytes): 0
	Average total size (kbytes): 0
	Maximum resident set size (kbytes): 837976
	Average resident set size (kbytes): 0
	Major (requiring I/O) page faults: 3
	Minor (reclaiming a frame) page faults: 623453
	Voluntary context switches: 40483
	Involuntary context switches: 1085
	Swaps: 0
	File system inputs: 536
	File system outputs: 42736
	Socket messages sent: 0
	Socket messages received: 0
	Signals delivered: 0
	Page size (bytes): 4096
	Exit status: 0

and for the efficiency:
Good found tracks/event: 4, efficiency: 0.942444, fake-track rate: 0.0563714 clone rate 0.0184747

Summarising: we use more memory, we are 25% slower and the efficiencies are comparable (slightly lower efficiency with the PR, higher fakes and reduced clones)

[iouribelikov]

@mpuccio Thank you very much, Max for these important numbers !
Look, I would suggest merging these two PRs. So that we can move ahead with the code development.
( If it is OK with @shahor02. )

[shahor02]

OK for me once the tests passed. But, @mpuccio , is it clear why there is no improvement in the eff.?

[mpuccio]

@shahor02, not really. But I'm also not that convinced that we really lose 4% out of this effect: the efficiency was always saturating at 1 at high pt with high statistical accuracy (better than 4%). Currently CheckTracks.C puts at denominator particles that have at least a cluster in all layers, and it does it, if I am not mistaken, looking also across different ROFs, thus the inefficiency should have been already evident with the current dev.

[shahor02]

How much we can lose depends on the eventual strobe length (the better/shorter it is the more we can lose) and the signal strength (since the rise time depends on it). In any case, we need to evaluate the eff. as a function of the event distance to the strobe boundary. If I got it right, @iouribelikov tries to get it?

[iouribelikov]

@mpuccio The CheckTracks.C is probably not the best tool to study this effect. The natural variable would be the time difference between the boundaries of ALPIDE strobes and the moment an interaction happens, as was suggested by @shahor02.
I already gave it a try, and the results were not easy to understand (for me). Everything could still be OK, but I'd need to digest it.
Now, since this option is anyway optional, let's merge it. So that @mconcas could move on.

[mpuccio]

@shahor02 @davidrohr, I am not sure to understand why the fullCI is breaking on TrackITS and MCCompLabel while the normal o2 build (and my local build) works fine. Do you have any insights?

[shahor02]

@mpuccio HIP compilation complains on o2::dataformats::RangeRefComp not having default c-tor. It does have it but it is defined as GPUdDefault(). If I get it right, with HIP it will become __device__, i.e. will not be available on the host?
Do you try locally with HIP? Perhaps @davidrohr may suggest something?

[mpuccio]

Hi @shahor02, unless there are surprises in the latest CIs or any objections from you, I would merge the PR here as it is

[mpuccio]

CI fully green :)